Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking

This repo contains datasets and code files for Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking.

Table of Contents

• TrajImpute Dataset
• Trajectory Imputataion Benchmarking
• Trajectory Prediction Benchmarking

---

TrajImpute Dataset

The dataset can be downloaded from Download Link. The structure of the TrajImpute dataset follows a dictionary format with specific keys as shown in /TrajImpute.png:

We use the same Dataloader for pedestrian trajectory generation as prior methods (referred from Social-GAN at https://github.com/agrimgupta92/sgan/blob/master/sgan/data/trajectories.py). Upon this data, we generate the missing observed trajectories. The missing value generation code is provided in data_generation.py.

python generate_data.py <data_test_file_path> <data_train_file_path> <data_val_file_path> <save_file_path>

---

Trajectory Imputataion Benchmarking

Results for Various Imputation Methods on Different Datasets

Results obtained for various imputation methods on the ETH-M, HOTEL-M, UNIV-M, ZARA1-M, and ZARA2-M subsets of TrajImpute with the easy protocol ($0 \leq \text{missing} \leq 4$) and the hard protocol ($4 \leq \text{missing} \leq 7$). The reported results show that SITES performs relatively better when imputing missing values. `M' refers to missing, indicating that the subset contains missing observed coordinates.

• Code for the Transformer [Code .]
• Code for the US-GAN [Code]
• Code for the BRITS [Code]
• Code for the M-RNN [Code]
• Code for the TimesNet [Code]
• Code for the SAITS [Code]

Datasets	Methods	Metrics	Transformer	US-GAN	BRITS	M-RNN	TimesNet	SAITS
ETH-M	Easy-impute	MAE	3.1318	0.6467	1.4287	5.2558	1.1353	0.5031
		MSE	19.4576	1.8055	4.7339	35.3738	4.9441	0.9909
		RMSE	4.4111	1.3437	2.1758	5.9476	2.2235	0.9954
		MRE	0.5236	0.1081	0.2389	0.8787	0.1898	0.0841
ETH-M	Hard-impute	MAE	3.2249	3.0451	3.0371	5.3309	1.3656	0.9965
		MSE	19.5948	18.0716	17.9457	35.5047	4.9937	2.5934
		RMSE	4.7926	4.2511	4.2362	5.9965	2.5054	1.6104
		MRE	0.5734	0.5100	0.5087	0.8962	0.2287	0.1669
HOTEL-M	Easy-impute	MAE	8.8847	2.6327	3.9033	3.2133	7.4037	2.1930
		MSE	91.5550	13.5993	23.1058	20.0857	124.5438	8.7460
		RMSE	9.5684	3.6877	4.8068	4.4817	11.1599	2.9574
		MRE	2.9468	0.8732	1.2946	1.0658	2.4556	0.7274
HOTEL-M	Hard-impute	MAE	8.9096	7.8833	7.6057	3.2443	7.9484	2.6050
		MSE	92.2607	75.9804	72.0169	20.2543	106.7010	16.0168
		RMSE	9.6478	8.7167	8.4863	4.5005	11.3296	4.0021
		MRE	2.8866	2.6127	2.5207	1.1686	2.6343	0.8634
UNIV-M	Easy-impute	MAE	3.0410	0.9158	1.0171	6.8380	0.6713	0.1939
		MSE	14.0163	2.6297	2.9769	56.9715	0.7631	0.0697
		RMSE	3.7438	1.6216	1.7254	7.5479	0.8736	0.2639
		MRE	0.3905	0.1176	0.1306	0.8780	0.0862	0.0249
UNIV-M	Hard-impute	MAE	3.9795	1.9430	1.8028	6.9148	0.9421	0.6158
		MSE	15.4244	6.1815	5.4057	57.6533	1.5827	0.6003
		RMSE	3.9639	2.4863	2.3250	7.7268	1.2581	0.7748
		MRE	1.0326	0.2495	0.2315	0.9751	0.1210	0.0791
ZARA1-M	Easy-impute	MAE	2.6288	0.4832	0.7307	5.1152	0.3125	0.2054
		MSE	10.0109	0.8599	1.2306	34.9869	0.1768	0.0775
		RMSE	3.1640	0.9273	1.1093	5.9150	0.4204	0.2784
		MRE	0.4326	0.0795	0.1202	0.8417	0.0514	0.0338
ZARA1-M	Hard-impute	MAE	2.7532	2.2846	2.3140	5.1921	0.5699	0.6277
		MSE	10.1228	7.8216	8.0351	35.7821	0.6327	0.8287
		RMSE	3.1816	2.7967	2.8346	5.9976	0.7955	0.9103
		MRE	0.4463	0.3756	0.3805	0.8673	0.0937	0.1032
ZARA2-M	Easy-impute	MAE	2.1301	0.3861	0.5556	5.0905	0.2409	0.1314
		MSE	7.3276	0.6212	0.8292	31.5674	0.1329	0.0385
		RMSE	2.7070	0.7882	0.9106	5.6185	0.3645	0.1963
		MRE	0.3524	0.0639	0.0919	0.8422	0.0399	0.0217
ZARA2-M	Hard-impute	MAE	2.2840	1.8605	1.8051	5.1698	0.5031	0.3632
		MSE	7.6342	5.8511	5.5953	32.3531	0.6525	0.4313
		RMSE	2.8630	2.4189	2.3654	5.8994	0.8077	0.6567
		MRE	0.3735	0.3041	0.2951	0.8465	0.0823	0.0593

---

Trajectory Prediction Benchmarking

Results obtained for various trajectory prediction methods on the imputed subsets of TrajImpute

We report the ADE/FDE for the trajectory prediction task on the clean, soft imputed, and hard imputed protocols. `Clean' refers to a subset with no missing coordinates. Performance degradation occurs when trajectory prediction is performed on the hard imputed subsets.

• Code for the GraphTern [Code]
• Code for the LBEBM-ET [Code]
• Code for the SGCN-ET [Code]
• Code for the EQmotion [Code]
• Code for the TUTR [Code]
• Code for the GPGraph [Code]

Datasets	Baselines	GraphTern	LBEBM-ET	SGCN-ET	EQmotion	TUTR	GPGraph
ETH	Clean	0.42/0.58	0.36/0.53	0.36/0.57	0.40/0.61	0.40/0.61	0.43/0.63
	Easy-impute	0.77/0.74	0.37/0.55	0.42/0.71	0.46/0.62	0.54/0.73	0.45/0.75
	Hard-impute	0.78/0.77	0.85/1.07	1.07/1.44	0.47/0.63	1.12/1.53	0.92/0.93
Hotel	Clean	0.14/0.23	0.12/0.19	0.13/0.21	0.12/0.18	0.11/0.18	0.18/0.30
	Easy-impute	0.15/0.25	0.13/0.20	0.14/0.23	0.65/0.68	1.31/1.66	0.19/0.31
	Hard-impute	1.68/1.42	3.31/4.13	3.21/3.92	0.72/0.74	3.36/3.95	1.89/1.70
UNV	Clean	0.26/0.45	0.24/0.43	0.24/0.43	0.23/0.43	0.23/0.42	0.24/0.42
	Easy-impute	0.27/0.47	0.30/0.51	0.29/0.51	0.37/0.61	0.31/0.49	0.25/0.44
	Hard-impute	0.50/0.51	0.64/1.01	0.77/1.21	0.39/0.70	0.59/0.85	0.53/0.50
ZARA1	Clean	0.21/0.37	0.19/0.33	0.20/0.35	0.18/0.32	0.18/0.34	0.17/0.31
	Easy-impute	0.22/0.38	0.20/0.35	0.22/0.38	0.27/0.43	0.24/0.41	0.18/0.32
	Hard-impute	0.96/1.25	0.37/0.60	0.61/0.97	0.28/0.44	0.50/0.77	0.58/0.45
ZARA2	Clean	0.17/0.29	0.14/0.24	0.15/0.26	0.13/0.23	0.13/0.25	0.15/0.29
	Easy-impute	0.18/0.30	0.16/0.27	0.17/0.29	0.36/0.54	0.25/0.37	0.29/0.30
	Hard-impute	0.37/0.44	0.27/0.43	0.41/0.63	0.37/0.55	0.33/0.50	0.36/0.34